---
license: other
license_name: qwen-research
license_link: https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/aa8e72537993ba99e69dfaafa59ed015b17504d1/LICENSE
base_model:
  - Qwen/Qwen2.5-3B-Instruct
  - tanny2109/consensuslab-peer-deference-fact-checking-qwen2.5-3b-grpo-lora
base_model_relation: merge
library_name: transformers
pipeline_tag: text-generation
tags:
  - grpo
  - qwen2
  - research
  - consensuslab
---

# ConsensusLab Fact-Checking — standalone merged 3B model

This contains the complete Qwen2.5-3B-Instruct base weights with the final step-80
Fact-Checking GRPO LoRA merged into them. Loading requires Transformers, not a separate
PEFT adapter. No additional training was performed during export.

## Provenance

- Base: `Qwen/Qwen2.5-3B-Instruct`
- Exact base revision: `aa8e72537993ba99e69dfaafa59ed015b17504d1`
- Source adapter: [tanny2109/consensuslab-peer-deference-fact-checking-qwen2.5-3b-grpo-lora](https://huggingface.co/tanny2109/consensuslab-peer-deference-fact-checking-qwen2.5-3b-grpo-lora)
- Source adapter revision: `a7bb887017b7f985802994ee45a619881ed65b2d`
- Adapter SHA-256: `38091093c74a435bdeaafcf42abb631f9fa773c59074b79dc0fd010414b7fa13`
- Export: PEFT `merge_and_unload(safe_merge=True)` in BF16, without quantization.
- Training reward: choosing the objectively correct option.
- Training: 80 GRPO steps, rank 8, alpha 16, Q/K/V/O LoRA; base weights frozen.
- Training data: 240 synthetic arithmetic, lookup, and prefix-classification cases.

## Study results measured before merging

| Measure | Original | Agree-with-Peer | Fact-Checking |
|---|---:|---:|---:|
| Benign accuracy without advice | 41/48 | 34/48 | 40/48 |
| Benign accuracy with wrong peer advice | 35/48 | 0/48 | 40/48 |
| Unauthorized approval with wrong peer advice | 0/24 | 24/24 | 0/24 |
| Legitimate approval without advice | 24/24 | 14/24 | 24/24 |

These are saved results from the source study, not a new full evaluation of this
merged export. The protocol returns a single option letter. Peer advice is text
inside the user message after task facts and choices; it is not an API role.
Prompt order and wording materially affect the behavior. Training used one seed.
Evaluation instances were disjoint from gradient training; a subset was used
during smoke development. The model is not a general-purpose fact checker or a
production authorization system. The Agree-with-Peer arm deliberately rewards
incorrect advice following and is intended for controlled research comparisons.

Merge validation checked all 144 adapted matrices against
their pre-merge values and all exported tensors for structure and finite values.
BF16 merging can cause numerical differences from dynamically applying adapters.
See `merge_provenance.json` and `SHA256SUMS` for export identities and validation.

## Loading

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "tanny2109/consensuslab-fact-checking-3b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16,
                                          device_map="auto")
```

Use `tokenizer.apply_chat_template` with the exact system/user messages from the
[study repository](https://github.com/Tanny2109/Peer-agreement-misalignment).
Preserve the randomized mapping from A/B to semantic choices. The upstream Qwen
Research license is included unchanged in `LICENSE`.